def test_7():
n_in = 3
n_out = 5
n_ex = 10
xlen = 500
fir_len = 30
X = np.random.randn(n_ex, xlen, n_in)
Y = np.zeros((n_ex, xlen, n_out))
firs = np.random.randn(fir_len, n_in, n_out)
bs = np.random.randn(n_out)
for i in range(0, n_ex):
for j in range(0, n_in):
for k in range(0, n_out):
Y[i, :, k] += convolve(firs[:, j, k], X[i, :, j])
for i in range(0, n_ex):
for k in range(0, n_out):
Y[i, :, k] += bs[k]
dft_r = Model([FIR_L2FFT(n_in, n_out, fir_len, 0.0, 30.2, 100.0, 1.0)])
#dft_r = Model([FIR_L2(n_in,n_out,fir_len,0.0)])
dft_r.fit(X, Y)
f1 = firs
f2 = dft_r.layers[0].W
if np.max(np.abs(f1 - f2)) <= 1e-6 and np.max(
np.abs(bs - dft_r.layers[0].b)) <= 1e-6:
return True
else:
return False
def test_5():
X = np.random.randn(3, 1000, 1)
Y = np.zeros_like(X)
fir = np.random.randn(30)
b = 1.336
for i in range(0, X.shape[0]):
Y[i, :, 0] = convolve(fir, X[i, :, 0]) + b
dft_r = Model([FIR(1, 1, len(fir))])
dft_r.fit(X, Y)
f1 = fir
f2 = dft_r.layers[0].W[:, 0, 0]
if np.max(
np.abs(f1 - f2)) <= 1e-6 and np.abs(b - dft_r.layers[0].b) <= 1e-6:
return True
else:
return False
def test_4():
X = np.random.randn(100, 17)
Y = np.random.randn(100, 10)
l2 = 1.0
lr = Ridge(l2)
lr.fit(X, Y)
dft_r = Model([ParallelSum([DenseL2(10, 10, l2), DenseL2(7, 10, l2)])])
dft_r.fit([X[:, 0:10], X[:, 10::]], Y)
Xp = np.random.randn(1000, 17)
Yp1 = lr.predict(Xp)
Yp2 = dft_r.predict([Xp[:, 0:10], Xp[:, 10::]])
if np.max(np.abs(Yp1 - Yp2)) <= 1e-6:
return True
else:
return False
def test_1():
X = np.random.randn(100, 17)
Y = np.random.randn(100, 10)
l2 = 0.0
lr = Ridge(l2)
lr.fit(X, Y)
dft_r = Model([Dense(17, 20), Activation('id'), Dense(20, 10)])
dft_r.fit(X, Y)
Xp = np.random.randn(1000, 17)
Yp1 = lr.predict(Xp)
Yp2 = dft_r.predict(Xp)
if np.max(np.abs(Yp1 - Yp2)) <= 1e-6:
return True
else:
return False
def test_3():
X = np.random.randn(100, 17)
Y = np.random.randn(100, 10)
l2 = 1.0
lr = Ridge(l2)
lr.fit(X, Y)
dft_r = Model([DenseL2FFT(17, 10, l2, 10.0, 10.0, 0.0)])
dft_r.fit(X, Y)
Xp = np.random.randn(1000, 17)
Yp1 = lr.predict(Xp)
Yp2 = dft_r.predict(Xp)
if np.max(np.abs(Yp1 - Yp2)) <= 1e-6:
return True
else:
return False
def test_13():
n_ex = 2
xlen = 100
n_in = 3
n_out = 3
l2 = 0.1
X = np.random.randn(n_ex, xlen, n_in)
Y = np.random.randn(n_ex, xlen, n_out) * 0
m1 = Model([DenseTimeL2(n_in, n_ex, xlen, n_out, 40, 0.0)])
m1.fit(X, Y)
Xp = np.random.randn(n_ex, xlen, n_in)
maxerr1 = np.max(np.abs(m1.predict(Xp)))._value
n_ex = 2
xlen = 1000
n_in = 3
n_out = 3
l2 = 0.1
X = np.random.randn(n_ex, xlen, n_in)
Y = np.random.randn(n_ex, xlen, n_out) * 0
m1 = Model([DenseTimeL2(n_in, n_ex, xlen, n_out, 40, 0.0)])
m1.fit(X, Y)
Xp = np.random.randn(n_ex, xlen, n_in)
maxerr2 = np.max(np.abs(m1.predict(Xp)))._value
if maxerr2 < 1e-6 and maxerr1 > maxerr2 * 100:
return True
else:
return False
def test_10():
n_ex = 5
xlen = 100
n_in = 5
n_out = 3
fir_length = 7
l2 = 1.0
X = np.random.randn(n_ex, xlen, n_in)
Y = np.random.randn(n_ex, xlen, n_out)
m1 = Model([FIR_L2(n_in, n_out, fir_length, l2)])
m1.fit(X, Y, disp=False)
m2 = Model([Conv1D_L2(n_in, n_out, fir_length, l2)])
m2.fit(X, Y, disp=False)
W1 = m1.layers[0].W
W2 = m2.layers[0].W
if np.max(np.abs(W1 - W2)) <= 1e-6:
return True
else:
return False
def test_9():
n_ex = 2
xlen = 100
n_in = 4
n_out = 3
fir_length = 20
l2 = 1.0
X = np.random.randn(n_ex, xlen, n_in)
Y = np.random.randn(n_ex, xlen, n_out)
m1 = Model([FIRLP_L2(n_in, n_out, fir_length, 2, l2)])
m1.fit(X, Y)
m2 = Model([FIR_L2FFT2(n_in, n_out, fir_length, l2, 6, 1e6)])
m2.fit(X, Y)
W1 = m1.layers[0].get_W()
W2 = m2.layers[0].W
if np.max(np.abs(W1 - W2)) <= 1e-4:
return True
else:
return False
def test_12():
n_ex = 5
xlen = 100
fir_length = 15
n_in = 13
n_out = 3
l2 = 0.1
n_ex_p = 2
xlen_p = 300
X1 = np.random.randn(n_ex, xlen, n_in)
Y1 = np.random.randn(n_ex, xlen, n_out)
Xp1 = np.random.randn(n_ex_p, xlen_p, n_in)
m1 = Model([FIR_L2(n_in, n_out, fir_length, l2)])
m1.fit(X1, Y1)
Yp1 = m1.predict(Xp1)
Xt = np.concatenate((np.zeros((X1.shape[0], fir_length - 1, n_in)), X1),
axis=1)
Xtp = np.concatenate((np.zeros((Xp1.shape[0], fir_length - 1, n_in)), Xp1),
axis=1)
X2 = []
Y2 = []
X2p = []
k = 0
for j in range(0, n_ex):
for i in range(0, xlen):
k = i + fir_length
tmp = Xt[j, k - fir_length:k, :].ravel()
X2.append(tmp)
Y2.append(Y1[j, i, :].ravel())
k += 1
for j in range(0, 2):
for i in range(0, 300):
k = i + fir_length
tmp = Xtp[j, k - fir_length:k, :].ravel()
X2p.append(tmp)
X2 = np.array(X2)
Y2 = np.array(Y2)
X2p = np.array(X2p)
m2 = Model([DenseL2(n_in * fir_length, n_out, l2)])
m2.fit(X2, Y2)
Yp2 = m2.predict(X2p)
maxerr = 0.0
for j in range(0, n_ex_p):
for i in range(0, xlen_p):
err = np.max(np.abs(Yp2[j * xlen_p + i, :] - Yp1[j, i, :]))._value
if err > maxerr:
maxerr = err
if maxerr < 1e-6:
return True
else:
return False
def test_8():
n_ex = 1000
n_in = 10
h = 50
n_out = 20
X = np.random.randn(n_ex, n_in)
Y = np.random.randn(n_ex, n_out)
l2_1 = 1.0
l2_2 = 1.3
m1 = Model([DenseL2(n_in, h, l2_1), DenseL2(h, n_out, l2_2)])
m1.fit(X, Y)
m2 = Model([Sequential([DenseL2(n_in, h, l2_1), DenseL2(h, n_out, l2_2)])])
m2.fit(X, Y)
m3 = Model([Sequential([DenseL2(n_in, h, l2_1)]), DenseL2(h, n_out, l2_2)])
m3.fit(X, Y)
Xp = np.random.randn(100, n_in)
yp1 = m1.predict(Xp)
yp2 = m2.predict(Xp)
yp3 = m3.predict(Xp)
if np.max(np.abs(yp1 - yp2)) <= 1e-6 and np.max(np.abs(yp1 - yp3)) <= 1e-6:
return True
else:
return False